Pattern learning and recognition apparatus in a computer system

• US 5,040,214 A
• Filed: 03/08/1989
• Issued: 08/13/1991
• Est. Priority Date: 11/27/1985
• Status: Expired due to Fees

• Current Assignee: Boston University
• Original Assignee: Boston University
• Inventors: Cohen, Michael, Grossberg, Stephen
• Primary Examiner(s): NOT, DEFINED
• Assistant Examiner(s): Merecki, John A.
• Application Number:
  US07/320,806
• Time in Patent Office: 888 Days
• Field of Search: 381/41, 381/47, 364/513, 364/513.5, 364/200 MS File, 364/900 MS File, 364/807, 382/14-15, 307/201
• US Class Current: 704/240
• CPC Class Codes: G06V 10/451 with interaction between the filter responses, e.g. cortical complex cells
• CPC Class Codes: G10L 15/16 using artificial neural networks

Abstract

A masking field network F₂, is characterized through systematic computer simulations serves or a content addressable memory. Masking field network F₂ receives input patterns from an adaptive filter F₁ →F₂ that is activated by a prior processing level F₁. The network F₂ activates compressed recognition close that are predictive with respect to the activation patterns flickering across F₁, and competitively inhibits, or masks, codes which are unpredictive with respect to the F₁ patterns. The masking field can simultaneously detect multiple groupings within its input patterns and assign activation weights to the recognition codes for these groupings which are predictive with respect to the contextual information embedded within the patterns and the prior learning of the network. Automatic rescaling of sensitivity of the masking field as the overall size of an input pattern changes, allows stronger activation of a code for the whole F₁ pattern than for its salient parts. Network F₂ also exhibits adaptive sharpening such that repetition of a familiar F₁ pattern can tune the adaptive filter to elicit a more focal spatial activation of its F₂ recognition code than does an unfamiliar input pattern. The F₂ recognition code also becomes less distributed when an input pattern contains more contextual information on which to base an unambiguous prediction of the F₁ pattern being processed. Thus the masking field embodies a real-time code to process the predictive evidence contained within its input patterns. Such capabilities are useful in speech recognition, visual object recognition, and cognitive information processin

GOVERNMENT SUPPORT

This invention was made with Government support under AFOSR-85-0149 awarded by the Air Force. The Government has certain rights in this invention.

Claims

1. Pattern recognition apparatus in a computer system, the apparatus comprising a feedback, self compensating network, the network receiving an input pattern provided to the computer system for recognizing and determining subpatterns thereof, the network simultaneously coding, through direct access of a content-addressable memory area, both the whole input pattern and various groupings of subpatterns in the input pattern, each said coding including a respective activity weight, the weight of a code indicating the probability of the input pattern being the grouping of subpatterns of that code, the subpatterns being independently recognizable patterns having exiting nodes in the memory area, the probability based upon spatial likeness between the grouping of subpatterns and the input pattern, with respect to previous and succeeding input patterns and past probabilities used for recognizing other input patterns, such that the code with a weight of highest probability indicates recognition of the input pattern as the grouping of subpatterns of the code.

2. Apparatus as claimed in claim 1 wherein the input patterns are a time ordered series of speech elements and the network coding includes context-sensitive language representations of the series of speech elements;

and the network further comprising computer means responsive to said representations of the series of speech elements, the computer means grouping the speech elements in a manner which reorganizes past groupings of the speech elements for coding through the content-addressable memory area, the computer means grouping the speech elements as a function of the context-sensitive language representations of the series of speech elements, such that the speech elements themselves and temporal order thereof are used together in the coding through the content-addressable memory area.

3. Apparatus as claimed in claim 1 wherein the network exhibits simple rules of neuronal development including (i) random increase of memory area access for input patterns and subpatterns, (ii) expansion of the memory area as a function of said increase of memory area access, and (iii) competition for access of particular nodes of the memory area, the development rules generating a network architecture whose simultaneous coding can directly activate correct subpattern groupings without the need for prior search through the memory area.

4. Apparatus as claimed in claim 1 wherein the network accomplishes direct access by performing multiple spatial scale analysis of temporally evolving input patterns which enhances correct encoding and competitively masks inappropriate encodings.

5. Apparatus as claimed in claim 1 wherein memory access provides a code which represents the current input pattern, and wherein this code indicates a probability of subsequent changes to the current input pattern by assigning activity weights indicative of activities to groupings which have not yet fully occurred, based on coding of previous input patterns and context including similarities between the previous input patterns and the current input pattern, without serial processing within the network.

6. A computerized pattern recognition and learning system comprising:

an adaptive filter that is activated by a source pattern and provides an input pattern indicative of the source pattern;

a self-similar, automatically gain controlled cooperative-competitive nonlinear feedback masking field network having a content addressable memory responsive to the adaptive filter input pattern, the content addressable memory holding a plurality of list codes, and based on the input pattern, the network activating list codes that are predictive of the source pattern and competitively inhibiting list codes which are unpredictive of the source patterns based on different groupings of subpatterns in the input pattern, such that the masking field network provides the list code in the content addressable memory which is most predictive of the source pattern to indicate recognition of the source pattern and stores the source pattern with the list code in the content addressable memory to provide learning in the system.

7. A pattern recognition and learning system as claimed in claim 6 wherein the masking field network further simultaneously detects multiple groupings of subpatterns within the input pattern and assigns activation weights to the list codes for these groupings, said weights based on contextual information embedded within the input pattern and on list codes activated for recognizing other previous input patterns, the subpatterns each having a respective list code in the content addressable memory, and the masking field network determining said contextual information from relative positions of the subpatterns in the input pattern.

8. A pattern recognition and learning system as claimed in claim 6 wherein the content addressable memory activates predictive list codes and inhibits unpredictive list codes through a multiple scale analysis.

9. A pattern recognition and learning system as claimed in claim 8 wherein recognition by the system of other previous input patterns biases the adaptive filter to signal the masking field network to activate preferred list codes.

10. A pattern encoding system as claimed in claim 9 wherein the prior learning of the system biases the adaptive filter through feedback means.

11. A pattern recognition and learning system as claimed in claim 7 wherein the masking field network has sensitivity to the input pattern and automatically rescales said sensitivity as the overall size of the input pattern changes but remains sensitive to salient subpatterns of the input pattern, such that the masking field network activates the list code at a different level for the whole input pattern than the list codes for salient subpatterns of the input pattern, and said masking field network activates the list code for a subpattern at a different level when the subpattern becomes a pattern whole in a new source context.

12. A pattern recognition and learning system as claimed in claim 7 wherein the masking field network is primed by inputs directly from the source pattern such that the masking field network activates list codes which represent predictions of how the source pattern may evolve in a subsequent time interval.

13. A pattern recognition and learning system as claimed in claim 7 wherein the masking field network exhibits adaptive sharpening in which repetition of a familiar source pattern tunes the adaptive filter to elicit from the content addressable memory a more focal spatial activation of list codes than does an unfamiliar source pattern.

14. A pattern recognition and learning system as claimed in claim 7 wherein the adaptive filter is tuned to elicit activation of preferred list codes which are held in a short term memory of the system.

15. A pattern recognition and learning system as claimed in claim 7 wherein the list codes activated become less distributed when the input pattern contains more contextual information on which to base an unambiguous prediction of the source pattern being processed.

16. A pattern recognition and learning system as claimed in claim 7 further comprising a multiplicity of source nodes for holding the source pattern the adaptive filter being responsive to the source nodes to provide an input pattern indicative of the source pattern.

17. A pattern recognition and learning system as claimed in claim 16 wherein the adaptive filter is differently responsive to different source nodes.

18. A pattern recognition and learning system as claimed in claim 7 wherein the content addressable memory includes a plurality of masking field nodes and holds the list codes in respective masking field nodes, the nodes being activity-dependent and self-similarly established to hold a source pattern with a corresponding list code.

19. A pattern recognition and learning system as claimed in claim 18 wherein a relative value of an activity dependent self-similarly established node is constrained by number of times the network activates the node for an adaptive filter input pattern.

20. A computerized recognition and learning system comprising:

a plurality of source nodes which embody a source pattern;

a multiple scale, cooperative-competitive feedback masking field network of list nodes, the list nodes holding respective recognition codes of patterns;

a plurality of adaptive filter paths which map the source nodes to list nodes in a manner which provides an input pattern to the network;

the network receiving the input pattern from the adaptive filter paths and having content-addressable memory means responsive to the input pattern, the content-addressable memory means determining (i) list nodes which provide compressed recognition codes that are predictive of the source pattern, said list nodes being activated list nodes, and (ii) list nodes which provide compressed recognition codes that are unpredictive of the source pattern, said list nodes being masked list nodes, where at least one adaptive filter path is mapped from each source node to at least one activated list node, such that the masking field network provides the recognition code which is most predictive of the source pattern, based on different groupings of subpatterns in the input pattern, to indicate recognition of the source pattern, and stores the recognition code of the source pattern in a list node to provide learning in the system.

21. A recognition and learning system as claimed in claim 20 wherein the adaptive filter paths are mapped in a profuse random manner.

22. A pattern recognition and learning system as claimed in claim 21 wherein the randomness of mapping from a source node is controlled by the number of adaptive filter paths associated with that source node.

23. A recognition and learning system as claimed in claim 20 wherein the number of adaptive filter paths mapped to a list node determines the size of the list node, and the content-addressable memory means determines an activated list node through a subset of certain mapped paths depending at least on the number of paths mapped to the list node.

24. A recognition and learning system as claimed in claim 23 further comprising means for preventing a threshold number of adaptive filter paths from being mapped to an activated list node.

25. A recognition and learning system as claimed in claim 23 wherein the paths mapped to one activated list node experience a change in strength relative to each other as the one node changes in size.

26. A recognition and learning system as claimed in claim 20 wherein a list node exhibits activity dependent self-similar growth constrained by conservation of points of connection between the list node and filter paths.

27. A recognition and learning system as claimed in claim 20 wherein the list nodes are of different memory sizes, larger list nodes being able to hold recognition codes of longer patterns than smaller list nodes, and more adaptive filter paths able to be mapped to larger list nodes than smaller list nodes.

28. A recognition and learning system as claimed in claim 27 wherein different subsets of the adaptive filter paths provide different subsets of the input pattern which enables selective determination of activated list nodes of a certain interaction ability.

29. A recognition and learning system as claimed in claim 20 wherein the masking field network simultaneously detects multiple groupings of subpatterns in the input pattern and assigns activation weights to the recognition codes for these groupings, the weights being based on relative positions of the subpatterns in the input pattern and recognition codes used to predict prior source patterns of the system, the subpatterns having recognition codes provided by certain list nodes.

30. A recognition and learning system as claimed in claim 20 wherein the masking field network automatically rescales its response to the input pattern as the overall size of the input pattern changes but remains responsive to relative positions of subpatterns in the input pattern such that the content-addressable memory means differently determines an activated list node for producing the code for the whole input pattern than a list node for producing the code for salient subpatterns of the input pattern, and differently determines an activated list node for producing the code for a source pattern part when that part is a pattern whole in a new context.

31. A method of learning and recognizing patterns comprising the steps of:

providing an input pattern;

simultaneously detecting multiple groupings of subpatterns within the input pattern, each grouping being associated with a predefined code accessible through a content-addressable memory;

assigning respective weights to the codes for the detected multiple groupings;

activating the codes associated with the detected groupings such that the codes of the groupings with their respective weights each provide a probability that the input pattern is the grouping associated with that code according to relative positions of subpatterns in the input pattern and codes used to recognize prior input patterns; and

selecting the code with the weight of highest probability to indicate recognition of the input pattern as the grouping of subpatterns of the selected code.

32. A method as claimed in claim 31 further comprising the step of competitively inhibiting codes which are unpredictive of the input pattern.

33. A method as claimed in claim 32 wherein the step of activating codes and the step of competitively inhibiting codes include activating predictive codes and inhibiting unpredictive codes through a multiple scale analysis.

34. A method as claimed in claim 31 further comprising the step of automatically detecting new groupings of subpatterns as the overall size of the input pattern changes but remaining responsive to salient subpatterns of the input pattern, such that the codes are activated at a different level for the whole input pattern than the codes of the salient subpatterns of the input pattern, the code for a subpattern of the input pattern being activated at a different level when the subpattern becomes a pattern whole in a new input context.

35. A method as claimed in claim 31 further comprising the step of providing priming inputs directly from the input pattern such that codes which represent predictions of how the input pattern may evolve in a subsequent time interval are activated.

36. A method as claimed in claim 31 wherein the step of activating codes includes adaptively sharpening such that repetition of a familiar input pattern elicits a more focal spatial activation of codes than does an unfamiliar input pattern.

37. A method as claimed in claim 31 wherein the step of activating codes includes activating codes which are less distributed when the input pattern is similar to a recognized prior input pattern on which to base an unambiguous prediction of the input pattern being processed.